Prof. Jeehang Lee
September 24(Thu) - September 24(Thu), 2020
12PM
ZOOM (ID: 728-142-6028 )
ABC Seminar
Date: 12:00 PM, Thursday, September 24th
ZOOM 회의 참가 ID: 728-142-6028
Speaker: Prof. Jeehang Lee
Department of Human-Centered AI, Sangmyung Univ., South Korea
Title: Optimal control of human learning process using neuroscience-inspired AI.
Abstract:
Humans, organisms and autonomous robots share a common problem: how to survive in a complex and rapidly changing world with little prior knowledge. The key to this end is learning, and using what is learned to guide further decisions in as efficient and optimal way as possible. Accumulating evidence suggests that the human brain has the functional flexibility for effective learning and inference from few observations with ease to this end. This raises an optimistic expectation that neuro-computational understandings of how human brains implement reinforcement learning (RL) would offer useful clues to the design of AI that would be able to artificially boost performance in human learning. We propose a novel framework in which a neural network learns a behavioral task control policy to guide the human learning and inference process at neural levels. The proposed framework is based on a self-play neural network, a deep reinforcement learning algorithm operating in a latent state space of a computational model that was previously fitted to account for neural activity patterns of the ventrolateral prefrontal cortex, basal ganglia and hippocampus. By applying this paradigm to learning and inference tasks, we demonstrated that a deep RL can learn to control a human learning process in terms of both performance and efficiency. This finding implies that artificial intelligence technology can not only be used to perform tasks on behalf of humans but also change the way humans solve problems.